KIST's Neuromorphic AI Trains Six Times Faster Than Google's Method, With No Extra Hardware
Researchers at KIST (the Korea Institute of Science and Technology) have created A²SG, short for Adaptive and Asymmetric Surrogate Gradients. It trains spiking neural networks to world-class accuracy on ImageNet while requiring six times less compute than Google's method. The changes are software-only, so no hardware modifications are needed. The work was presented at ICML 2026 in Seoul.
AI-processed from TechXplore AI; edited by Hamidun News
On July 6, 2026, South Korea's KIST (Korea Institute of Science and Technology) presented A²SG (Adaptive and Asymmetric Surrogate Gradients) at ICML 2026. The technique trains spiking neural networks with roughly six times less compute than Google's leading method, at comparable accuracy.
What A²SG Is and Why It Matters
Spiking neural networks (SNNs) mimic the biological brain. Instead of passing a continuous stream of numbers, they transmit signals as sparse pulses, which gives them fundamentally lower power consumption than conventional deep neural networks (DNNs). The catch is that they are harder to train, because standard backpropagation does not work directly on SNNs.
A²SG was developed by a team led by senior researcher Songsik Park of KIST's Semiconductor Technology Division. It tackles the training problem with adaptive and asymmetric surrogate gradients for transformer-based SNN architectures.
- Accuracy: world-class results on the ImageNet benchmark
- Computational overhead: ~1/6 of Google's method
- Implementation: software-only, no hardware changes required
- Applications: smartphones, wearables, drones, 24-hour sensors
- Publication: arXiv (DOI: 10.48550/arxiv.2606.11236), ICML 2026, Seoul, July 6–11, 2026
What Changed in the Training Approach
The method's key difference from its predecessors is that it is both adaptive and asymmetric. Traditional surrogate gradients apply the same function to every neuron and in every direction of learning. A²SG tailors the shape of the gradient to each neuron based on its activity, and it uses different strategies for the forward and backward passes. This is what delivers the efficiency gain without a loss in quality.
"This research resolves structural problems in neuromorphic AI and raises its potential for practical use," said
Songsik Park, lead researcher at KIST.
According to arXiv, the method showed consistent improvements across networks of various sizes and on different tasks. The first author of the paper is graduate student Kang Yechan.
What This Means
Neuromorphic AI is moving out of its academic niche. If A²SG can be commercialized, and KIST has already announced plans to build low-power AI chips based on RPU (Random Processing Unit), AI-enabled devices could run without constant cloud support. They would draw fractions of a watt rather than the megawatts consumed by data centers.
Frequently Asked Questions
What is a spiking neural network, and how is it better than a conventional one?
A spiking neural network (SNN) conveys information through sparse pulses, like neurons in the brain, rather than a continuous stream of numbers. This cuts power consumption by orders of magnitude compared with classical DNNs, with comparable result quality on a number of tasks.
Do I need to change hardware to use A²SG?
No. The method is implemented entirely in software. There is no need to swap processors or chips, which sharply lowers the barrier to adoption.
- Oct 28, 2026A 2015 CNN Beat GPT-6 Astra and Claude Fable 5.1 at Skin Cancer Diagnosis
- Oct 28, 2026Google taught an LLM to “sleep”: SLEEP consolidates knowledge into weights, up to 10 million tokens
- Oct 28, 2026Keystone removes the model name from code: open-source LLM gateway adds 0.64 ms
- Oct 28, 2026Wired Tested 6 AI Agents in a Day: Dots Leads, Gemini Spark Fails at Lunch
Key Questions Answered
What is a spiking neural network, and how is it better than a conventional one?
A spiking neural network (SNN) conveys information through sparse pulses, like neurons in the brain, rather than a continuous stream of numbers. This cuts power consumption by orders of magnitude compared with classical DNNs, with comparable result quality on a number of tasks.
Do I need to change hardware to use A²SG?
No. The method is implemented entirely in software. There is no need to swap processors or chips, which sharply lowers the barrier to adoption.
What A²SG Is and Why It Matters?
Spiking neural networks (SNNs) mimic the biological brain. Instead of passing a continuous stream of numbers, they transmit signals as sparse pulses, which gives them fundamentally lower power consumption than conventional deep neural networks (DNNs). The catch is that they are harder to train, because standard backpropagation does not work directly on SNNs. A²SG was developed by a team led by senior researcher Songsik Park of KIST's Semiconductor Technology Division. It tackles the training problem with adaptive and asymmetric surrogate gradients for transformer-based SNN
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.